Character consistency is one of the hardest parts of AI video creation. A prompt can describe the same person in every shot and still produce a different face, outfit, silhouette, or product detail. The solution is not to add more adjectives. It is to give the model a clearer visual anchor and reduce the number of things that can change at once.
Veo3o1 provides two image-based workflows for this job: Reference Images and First & Last Frames. Reference Images are usually the better starting point when identity matters more than an exact opening or ending composition.
What reference images control
A reference image gives the model visual information that is difficult to describe precisely in text: facial proportions, clothing, color palette, packaging, materials, or a recurring object. In Veo3o1, Reference Images mode accepts one to three images and is available with the Lite and Fast model presets.
Reference mode currently uses:
- one to three reference images;
- Lite or Fast;
- an 8-second duration;
- a fixed 16:9 or 9:16 aspect ratio rather than Auto.
Those constraints are visible before generation. The exact credit cost is also shown before you submit. See the Reference Images documentation for the current control rules.
Build a useful reference set
More images are not automatically better. Each image should answer a different visual question.
For a character, a practical three-image set is:
- a clear front or three-quarter portrait;
- a wider image showing clothing and body proportions;
- a side view or a second expression under similar lighting.
For a product, use:
- a clean hero angle;
- an angle that reveals shape and materials;
- a detail view of the most important feature.
Avoid mixing different outfits, hairstyles, package versions, color grades, or lighting directions unless the change is intentional. Conflicting references force the model to guess which detail should win.
Write a prompt that protects identity
Treat identity as fixed and motion as variable. A useful structure is:
[fixed subject] + [single action] + [environment] + [camera behavior] + [lighting] + [details that must remain unchanged]
Example:
The woman shown in the reference images walks slowly through a quiet hotel lobby, keeping the same short black hair, navy suit, silver earrings, and facial features. Medium tracking shot at eye level, warm window light, restrained natural motion, no wardrobe change.
The prompt does four jobs. It identifies the reference subject, asks for one readable action, defines the shot, and repeats the details that matter. It does not introduce a second character, a costume change, a time jump, and a complex camera move in the same eight seconds.
Use controlled iterations
Change one variable per generation. Start with the subject, action, and camera. If the identity is stable, refine lighting or background in the next attempt. If you change the prompt, seed, aspect ratio, and model together, you will not know which change helped.
An optional seed can make repeated settings easier to compare, but it is not a guarantee that every frame will be identical. Save the prompt, selected model, aspect ratio, reference set, and seed with each useful result.
Diagnose common failures
The face or product shape drifts
Use a cleaner primary reference and simplify the action. Wide shots, rapid movement, heavy motion blur, and long occlusion give the model less visible identity information.
Clothing or colors change
Remove conflicting references. Name the fixed color and garment once in a direct sentence. Avoid style words that imply a new palette.
The subject looks static
Keep the identity constraints, but describe one physical action and one camera action. “Turns toward the window while the camera makes a slow push-in” is easier to interpret than “make it cinematic and dynamic.”
The scene is overloaded
Divide the idea into separate clips. Use one shot to establish the subject, another for the action, and edit the clips together afterward. A short generation works best when it has one visual priority.
Reference Images or First & Last Frames?
Use Reference Images when you want to preserve the look of a person, product, or visual world while allowing the model to compose the motion. Use First & Last Frames when the exact starting image, ending image, or transition path is the main requirement. Our First & Last Frame comparison explains the decision in detail.